In DeepSeek's RL ablation on DeepSeek-R1-Distill-Qwen-7B, chain-of-thought language consistency declines without an LC reward but stays stable with one.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The DeepSeek-R1 paper (arXiv 2501.12948, Appendix B.6; later published in Nature) reports an ablation of its language consistency (LC) reward on DeepSeek-R1-Distill-Qwen-7B. With no LC reward, the share of chain-of-thought written in the target language falls steadily as RL training goes on. With the reward, it stays stable. The authors also report roughly unchanged math benchmark scores and a slight drop on coding. The experiment was not run on DeepSeek-R1-Zero, which was trained with no LC reward at all and showed language mixing.
The result is plausible, because an explicit reward for the target-language token share should hold that share up. It also fits the language mixing the paper and others saw in R1-Zero. Still, it is the developers' own single ablation, shown as one training curve, with no variance reported and no independent replication found.
Full reasoning: the evidence and decisions behind this verdict
Before searching: an LC reward defined as the proportion of target-language words in the CoT directly optimizes the measured quantity, so stability with the reward is close to built in. Without it, drift toward mixed language is consistent with the widely reported language mixing in R1-Zero. What could make the claim false is mainly misreporting or an artifact of a single run.
Evidence: Appendix B.6 of arxiv.org/pdf/2501.12948 states that the ablation is run on DeepSeek-R1-Distill-Qwen-7B. It reports, with Figure 7, that "without the LC reward, language consistency gradually deteriorates as training steps increase" and that with the reward "stable language consistency is maintained." The stored passage was checked verbatim. The original wording named DeepSeek-R1-Zero as the subject. That is a misattribution: R1-Zero is the run with no LC reward, so no source reports its consistency trajectory under one. The canonical wording now names the model actually tested.
Status is supported rather than verified. There is a single self-reported run shown only as a figure, and the underlying data were not examined. The credence reflects that the reported result is mechanistically expected and comes from a peer-reviewed (Nature) version of the paper.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
As can be seen, without the LC reward, language consistency gradually deteriorates as train- ing steps increase. However, when the LC reward is applied, stable language consistency is maintained throughout the training process.
Ablation on DeepSeek-R1-Distill-Qwen-7B in Appendix B.6.
The source's own evidence bears what it asserts. The ablation is run on DeepSeek-R1-Distill-Qwen-7B, not on DeepSeek-R1-Zero, and the result is shown as a single training-curve figure with no reported variance or repeated runs. The same section reports comparable math performance and slight coding degradation with the reward.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by extractor · Oct 28, 2026. Every judgment on this page is accompanied by a reasoning trace.